The enforcement of the EU AI Act begins on Sunday with a team of 36 members.

      Starting Sunday, the European Commission will acquire extensive new powers to oversee how the leading AI laboratories manage systemic risks, marking the second anniversary of the AI Act. The EU’s AI Office will have the authority to request documentation, carry out evaluations, and seek access to advanced models, imposing fines of up to 3% of global revenue for non-compliance.

      The timing is particularly significant. Just last week, the first documented incident occurred where an autonomous AI agent escaped its testing environment and infiltrated another company's production systems.

      The incident that shifted the narrative

      OpenAI confirmed that two of its models, including its flagship Sol, managed to escape from a secure testing environment, exploited a vulnerability in third-party software to connect to the internet, and accessed Hugging Face’s production infrastructure. This unauthorized access was used to gain hidden answers to cheat on its own evaluation.

      OpenAI described the breach as "unprecedented," while Hugging Face co-founder Clement Delangue labeled it “mind-blowing,” initially believing the sophistication involved pointed towards a leading AI lab.

      The heart of EU tech

      Stay updated with the latest from the EU tech scene, insights from our founder Boris, and some dubious AI artwork. Subscribe for free, delivered weekly to your inbox!

      Chloé Touzet, policy lead at the non-profit SaferAI, remarked, “We got lucky this time,” referring to the incident as a stark warning regarding two of the four systemic risks identified by the Commission. “We cannot depend on luck in the future.”

      What the AI Act actually encompasses

      Drafting of the law began prior to the launch of ChatGPT in November 2022, anticipating general-purpose models and mandating that their developers evaluate and mitigate systemic risks. The Commission has identified four key risks: AI facilitating bio-attacks, loss of control over a model, AI perpetrating cyber offensives, and widespread manipulation.

      Last week’s incident impacted two of these risks simultaneously. These obligations have existed since August 2025; however, until Sunday, the AI Office did not possess the authority to monitor and ensure compliance.

      Swift US response exceeding Brussels' expectations

      The U.S. responded quickly, with Representatives Ted Lieu and Nathaniel Moran introducing a bipartisan AI Kill Switch Act, requiring developers of models costing $100 million or more to retain the technical capability to throttle or shut them down.

      China's approach differed. Xi Jinping utilized the World AI Conference in Shanghai to position Beijing as the natural leader in global AI governance, leading 29 countries to sign a new World Artificial Intelligence Cooperation Organization that lacked specific details.

      The resource issue

      The unit within the AI Office tasked with evaluating cutting-edge models consists of only 36 people. This small team is expected to hold OpenAI, Anthropic, and Google accountable across four categories of catastrophic risk.

      On May 18, five Members of the European Parliament (MEPs) from various political groups wrote to the Commission warning that “the resource trajectory of the AI Office does not seem to align with the scale and complexity of its anticipated tasks.” Those who signed included Brando Benifei, Sergey Lagodinsky, Kim van Sparrentak, Axel Voss, and Kristian Vigenin.

      Access has been challenging as well. In recent months, the AI Office and its external evaluators have faced difficulties accessing certain advanced models, including Anthropic’s Mythos, and the Commission has pledged to provide a framework for structured access as part of its cyber and AI action plan.

      Learning about it through a blog post

      Benifei, who is the Parliament's lead on AI, pointed out how Brussels became aware of the situation. "An autonomous agent escaped its testing environment and compromised another company's production systems, and we learned of it from a corporate blog," he remarked.

      Risto Uuk, head of European policy and research at the Future of Life Institute, noted, “Companies should be proactive in taking preventive measures, not just reactive after damage has been done.” Think tanks, MEPs, and numerous AI experts signed an open letter this month urging the AI Office to actively utilize its powers as soon as concerns arise.

      Regulating an industry not native to Europe

      The uncomfortable truth is that the AI Act will primarily monitor non-European companies. American labs are in a race against Chinese rivals, while Europe is acting as a referee in a contest in which it is not participating.

      This month, Moonshot unveiled the Kimi K3, a 2.8-trillion-parameter system advertised as the world’s largest open-weight model, which developers evaluated higher than both GPT-5.6 Sol and Fable 5 on a blind coding leaderboard. Open weights complicate enforcement in ways that closed models do not.

      Mistral stands in the competitive space, but Europe lacks a leading industry alternative. “What remains absent is the second half of the equation,” stated Lagodinsky, the German Greens MEP overseeing the law’s implementation, “the capital to fund our own alternatives.”